Files
skyg-linux-tracker/plan/ws-h-final-fixes.md
T
2026-10-20 16:28:35 -04:00

135 lines
8.2 KiB
Markdown

# WS-H: Phase 5 final-review fixes
| Field | Value |
|---|---|
| Status | Done (awaiting merge) |
| Owner | goose |
| Owned files | `src/peers/**`, `src/telemetry/mod.rs`, `src/main.rs`, `tests/**`, `docs/consolidator.md`, `config.example.toml`, `README.md` |
## Log
### 2026-10-20 Handoff
The review text (`loop/final-review.md`) is not present in this worktree; the
six items below were taken from the task statement that quotes it. Nothing in
`contracts.md` was touched and no contract looked wrong.
All items are done, one commit each, `dev`-level gate green after every commit
(`cargo fmt --check`, `cargo clippy --all-targets -- -D warnings`,
`cargo test`, `cargo tree -i openssl` empty). No Nix files changed, so no nix
build or VM test was run.
1. **One token loader, one set of peer helpers** (`H: merge token loaders and
peer helpers`). `src/peers/client.rs` no longer has a loader: it calls
`secrets::load_token` for every request. The header-validity rule (no CR,
LF or NUL, and the value must be a valid `Bearer` header) lives in
`secrets.rs`. `secrets::under_credentials` canonicalises both paths and
fails closed, so the client's non-canonical `starts_with` is gone with
`load_token_in`. Deny-CIDR parsing (`config::deny_networks`), the loopback
test (`config::url_is_loopback`), the plain-http rule (`plain_http_allowed`,
`plain_http_off_loopback`) and its warning wording (`warn_plain_http`) exist
once in `config.rs` and are used by validation, the transport and
`registry::reload`. Tests: `secrets::tokens_that_cannot_be_a_header_value_are_refused`,
`client::a_refused_token_file_sends_nothing` (no request reaches the peer),
`config::deny_networks_rejects_a_bad_cidr_once_for_every_caller`,
`config::loopback_detection_covers_localhost_and_both_families`.
Deviation: the merged mode check is the stricter union of the two old ones
(`0o077`, message "group or world accessible" instead of the client's
"readable"), so a group/world-writable token file stays refused; wording
updated in `secrets.rs` and `docs/consolidator.md`. A doubled trailing
newline is now `invalid` (it would end the header line) instead of being
forwarded with one newline left.
2. **`denied()` canonicalises** (`H: canonicalise addresses in the deny check`).
`config::denied` calls `IpAddr::to_canonical`; config validation and the
transport both use it. Tests:
`config::ipv4_mapped_ipv6_literals_match_the_v4_deny_list` (unit) and
`client::an_ipv4_mapped_address_is_matched_against_the_v4_deny_list`
(`::ffff:169.254.169.254` refused before any connect); both fail if the
canonicalisation is removed.
3. **IPv6 deny defaults** (`H: deny IPv6 link-local and metadata addresses by
default`). `types::DEFAULT_DENY_CIDRS` adds `fe80::/10`, `::/128` and
`fd00:ec2::254/128`; synced in `config.example.toml`, `README.md`,
`docs/consolidator.md` and `tests/peers_config.rs` (`PEERS_DEFAULT`). Test:
`config::the_default_deny_list_covers_metadata_in_both_families_but_allows_lans`.
4. **`peer.request` is a client span** (`H: mark the peer.request span as an
OTel client span`). `otel.kind = "client"` on the span, asserted in
`telemetry::tests::peer_request_span_names_the_peer`.
5. **Gauge gating in the docs** (`H: correct the peer gauge gating in the
docs`). `docs/consolidator.md` now names `otel.enabled` with
`otel.export_unit_metrics` (there is no `otel.metrics.enabled`) and states
that `enabled = false` peers have no series; behaviour was already covered
by `metrics_export::tests::disabled_peer_is_not_reported`.
6. **Port-stealing flakiness** (`H: make the SIGTERM tests immune to a stolen
port`). `tests/tls.rs::try_start_on` and
`tests/integration.rs::sse_stream_then_sigterm_on` only accept a socket that
answers as this server (a TLS handshake against the test pair; a `200` from
`/healthz`) while the child is still running, and report "retry" otherwise;
`start`/the test then retry on a fresh port, up to three times. The SSE exit
bound is 10 s. New tests
`tests/tls.rs::a_probed_port_taken_by_someone_else_is_retried` and
`tests/integration.rs::a_probed_port_taken_by_someone_else_is_retried` fail
if the identity check is weakened back to a bare `connect()`.
### 2026-10-20 Handoff (review-1)
Every finding in `loop/H/review-1.md` is answered, one commit each. The full
gate is green after every commit: `cargo fmt --check`,
`cargo clippy --all-targets -- -D warnings`, `cargo test` (run twice; the
SIGTERM tests pass both times), `cargo tree -i openssl` empty. No Nix files
changed. `devenv shell -- dev-test all` could not be used in this worktree:
devenv aborts before any command with "To use 'git-hooks', run `devenv inputs
add git-hooks …`" (an environment problem, not a repo change), so the same
cargo commands were run directly.
1. **Blocking: the SSE SIGTERM test proves identity, not liveness**
(`H: prove the SSE test server identity`). `tests/integration.rs` spawns the
binary with stderr piped and waits for its own `listening addr=127.0.0.1:<port>`
line (`spawn_listening` / `logs_listening`, ANSI stripped, the address parsed
and compared whole so a neighbouring port cannot match). Only then is the
child signalled; `sigterm_and_wait` now also refuses to signal a child that
has already exited. A port a stranger won makes `spawn_listening` return
`None` and the test retries on a fresh port; the MCP `tail_journal` test
uses the same helper. `healthz_answers` is no longer an identity check.
New test `a_probed_port_held_by_a_lying_squatter_is_retried` starts a
thread that answers `HTTP/1.1 200` (verified by the test itself) and asserts
the helper retries; `only_the_childs_own_listening_line_counts` pins the
parser (other port, other address, non-startup line).
2. **Blocking: mode tests for every group/world bit**
(`H: pin the token mode check for every group/world bit`).
`secrets::group_or_world_write_and_execute_bits_are_rejected_too` covers
0620, 0602, 0610 and 0601, and
`tests/peers_config.rs::a_loose_token_fails_startup_and_per_request_checks_alike`
asserts that `validate_token_readiness` and `peers::load_token` give the
same reason, "group or world accessible", for each mode. Both fail with the
old `0o044` mask (checked by reverting the mask locally).
3. **The plain-HTTP warning no longer appears twice**
(`H: warn once about plain http`). The transport warning and its
`plain_warned` set are gone: every peer set is validated before the registry
uses it (startup and reload both go through `Config::load`), and
`plain_http_off_loopback_requires_the_flag_and_warns_when_set` now asserts
exactly one `cleartext` line. The remaining helper is
`config::warn_plain_http_peer`, named after the peer it reports.
Deviation: the duplicated warning survives only for a `PeersConfig` handed
to `PeerRegistry::reload` directly, bypassing validation; that is test-only
and `validate_target` still refuses a plain-http peer without the opt-in.
4. **A FIFO token file cannot block a thread**
(`H: never open a token path that is not a regular file`).
`load_token_at` calls `std::fs::metadata` and refuses non-files before
opening, then opens with `O_NONBLOCK` to close the window before the open,
and reads the mode and length from the open descriptor.
`a_fifo_is_refused_without_waiting_for_a_writer` (mkfifo, load on a thread,
5 s timeout) fails on the previous code with "must not wait on a fifo".
5. **`a_refused_token_file_sends_nothing` covers the whole loop**
(`H: keep the token listener open for the whole loop`). The listener is no
longer consumed by a 500 ms accept inside a task: it stays open for the
four cases and one bounded `accept()` after the loop must time out. Making a
case reachable (a 0600 token) fails the test.
6. **The span kind is checked where a backend sees it**
(`H: assert the exported peer request span kind`). A capturing
`SpanExporter` plus `SimpleSpanProcessor` exports the span;
`peer_request_span_is_exported_as_a_client_span` requires
`SpanKind::Client`, the name `peer.request`, and no exported `otel.kind`
attribute. It fails if the field is changed to `server`.
No contract looked wrong and `contracts.md` was not touched.